Journal of Environmental Management 95 (2012) 98e107

Contents lists available at SciVerse ScienceDirect

Journal of Environmental Management

journal homepage: www.elsevier.com/locate/jenvman

Remote sensing of aquatic vegetation distribution in Taihu Lake using an

improved classification tree with modified thresholds
Dehua Zhao*,1, Hao Jiang 1, Tangwu Yang, Ying Cai, Delin Xu, Shuqing An
State Key Laboratory of Pollution Control and Resource Reuse, Nanjing University, Nanjing 210093, PR China

a r t i c l e i n f o a b s t r a c t

Article history: Classification trees (CT) have been used successfully in the past to classify aquatic vegetation from
Received 18 February 2011 spectral indices (SI) obtained from remotely-sensed images. However, applying CT models developed for
Received in revised form certain image dates to other time periods within the same year or among different years can reduce the
27 September 2011
classification accuracy. In this study, we developed CT models with modified thresholds using extreme SI
Accepted 9 October 2011
values (CTm) to improve the stability of the models when applying them to different time periods. A total
Available online 10 November 2011
of 903 ground-truth samples were obtained in September of 2009 and 2010 and classified as emergent,
floating-leaf, or submerged vegetation or other cover types. Classification trees were developed for 2009
Keywords:
Aquatic vegetation
(Model-09) and 2010 (Model-10) using field samples and a combination of two images from winter and
Remote sensing summer. Overall accuracies of these models were 92.8% and 94.9%, respectively, which confirmed the
Classification tree ability of CT analysis to map aquatic vegetation in Taihu Lake. However, Model-10 had only 58.9e71.6%
Classification accuracy classification accuracy and 31.1e58.3% agreement (i.e., pixels classified the same in the two maps) for
aquatic vegetation when it was applied to image pairs from both a different time period in 2010 and
a similar time period in 2009. We developed a method to estimate the effects of extrinsic (EF) and
intrinsic (IF) factors on model uncertainty using Modis images. Results indicated that 71.1% of the
instability in classification between time periods was due to EF, which might include changes in
atmospheric conditions, sun-view angle and water quality. The remainder was due to IF, such as
phenological and growth status differences between time periods. The modified version of Model-10 (i.e.
CTm) performed better than traditional CT with different image dates. When applied to 2009 images, the
CTm version of Model-10 had very similar thresholds and performance as Model-09, with overall accu-
racies of 92.8% and 90.5% for Model-09 and the CTm version of Model-10, respectively. CTm decreased the
variability related to EF and IF and thereby improved the applicability of the models to different time
periods. In both practice and theory, our results suggested that CTm was more stable than traditional CT
models and could be used to map aquatic vegetation in time periods other than the one for which the
model was developed.
Ó 2011 Elsevier Ltd. All rights reserved.

1. Introduction
Aquatic macrophytes support many important ecological and
socioeconomic functions, such as stabilization of sediments, regu-
lation of the nutrient cycle, slowing of water currents and main-
tenance of fishery production (Orth and Moore, 1983; Orth et al.,
1984; Jackson et al., 2001). Influenced by degradation of the
water environment, aquatic macrophytes have suffered extensive
losses in the past decades (Gullström et al., 2006). Efficient and
accurate monitoring of the distribution and growth of macrophytes
* Corresponding author. Tel./fax: þ86 25 83592705.
E-mail address: dhzhao@nju.edu.cn (D. Zhao).
1
Co-first author.
is important for lake management (Diaz et al., 2004; Jones et al.,
2009). At large scales, conventional methods for conducting
macrophyte surveys can be impractical. Moreover, macrophytes
sometimes grow in difficult-to-reach places, which further adds to
the difficulties in mapping their distribution and monitoring their
growth.
Remote sensing can be an effective tool for mapping macro-
phyte distribution over large areas, although its use is more prob-
lematic for aquatic vegetation than for terrestrial vegetation (Vis
et al., 2003; Silva et al., 2008). During the past three decades,
many remote sensing techniques have been developed to automate
the identification of aquatic vegetation on remote sensing imagery,
including unsupervised isoclustering techniques, supervised
maximum likelihood classifiers, Tasseled-Cap classification and
remote sensing combined with ancillary environmental data (Work

0301-4797/$ e see front matter Ó 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jenvman.2011.10.007
 D. Zhao et al. / Journal of Environmental Management 95 (2012) 98e107
and Gilmer, 1976; Fyfe, 2003; Gullström et al., 2006; Dogan et al.,
2009). Additionally, non-parametric classifiers have been widely
implemented as an alternative to traditional remote sensing anal-
ysis techniques (Orth and Moore, 1983; Ozesmi and Bauer, 2002;
MacAlister and Mahaxay, 2009).
As a type of non-parametric classifier, classification trees (CT)
based on dichotomous partitioning have been used successfully to
classify aquatic and other vegetation types from remotely-sensed
images (Brown et al., 2003; Baker et al., 2006; Wright and
Gallant, 2007; Davranche et al., 2010). However, rule-based
methods such as CT have not always performed well in identi-
fying aquatic vegetation, especially submerged vegetation, because
of the disruptive influence of factors such as water turbidity and
morphological variation (Nelson et al., 2006; Wright and Gallant,
2007; Davranche et al., 2010).
The main factors influencing the performance of a classification
tree can be divided into two groups: intrinsic (i.e. aquatic vegeta-
tion conditions) and extrinsic (i.e. environmental or physical
conditions); this division follows the rationale of Graetz (1987),
who divided the characteristics of arid rangelands influencing the
application of remote sensing into intrinsic and extrinsic factors.
The intrinsic factors are determined primarily by the temporally
dynamic characteristics of aquatic vegetation. Phenological and
growth status characteristics of aquatic vegetation change with
time, resulting in related differences in spectral characteristics
among time periods. Therefore, using classification trees to identify
aquatic vegetation from spectral indices between time periods
necessarily involves much uncertainty and error. The extrinsic
factors affecting the performance of a classification tree relate
primarily to changes in environmental or physical conditions, such
as atmospheric conditions, water quality and sun-view angle,
which vary with time and thus greatly influence the remote sensing
signal (Nelson et al., 2006; Silva et al., 2008). As an extrinsic factor,
atmospheric condition is vitally important in determining the
quality of an image, especially in the wet season when atmospheric
water vapor is maintained at a high level (Tan and Li, 2000).
With enough training samples and images for the period of
interest, accurate classification trees can probably be developed.
However, the classification trees developed for one time period or
image date may not be adequate when applied to other periods or
dates, as they are influenced by the intrinsic and extrinsic factors.
Therefore, using classification trees with modified splitting
thresholds that incorporate the differing characteristics brought
about by intrinsic and extrinsic factors affecting the images of two
time periods or dates may prove to be a better strategy than using
traditional classification trees with rigid thresholds, especially for
submerged vegetation with reduced signal magnitude (Silva et al.,
2008; Bustamante et al., 2009). We define “splitting thresholds”
here as the threshold values of an independent variable (i.e.,
spectral index) that are used to segregate the individual samples
into two groups or “branches” in the classification tree, such that
samples with variable values higher than the splitting threshold
follow one branch and those with values lower than the splitting
threshold follow the other branch, under the premise that this
segregation creates two groups of samples that are more similar
within the group than between groups. To our knowledge, no
studies have been reported in the literature on the application of
classification trees with modified thresholds to identify aquatic
vegetation from remote sensing images.
Taihu Lake is the third-largest freshwater lake in China.
According to recent field surveys, aquatic vegetation distribution
was widest for submerged vegetation, followed by floating-leaf
vegetation and, finally, emergent vegetation; the distributions
and biomass were greatest at the end of summer for all of the
vegetation types (Yang, 1998; Liu et al., 2007a; He et al., 2008).
99
Field surveys have also indicated significant changes in aquatic
vegetation distribution during the past decades (Liu et al., 2007a).
Developing efficient, large-scale monitoring strategies is key to
effectively tracking and mitigating harmful changes in the lake.
The application of remote sensing technology to aquatic vege-
tation monitoring has special significance for lake management at
regional scales because of its potential for covering larger areas
than would be possible or feasible for field monitoring efforts. In
this study, we developed traditional classification tree models
using field data from 2009 to 2010 along with corresponding
Landsat ETM þ images and validated their ability to accurately map
the aquatic vegetation in similar and different time periods.
Subsequently, our primary objective was to develop a new classi-
fication tree method with modified thresholds to improve the
stability of classification trees applied to different time periods.
2. Materials and methods
2.1. Study area
Taihu Lake is shallow, with an average depth of 1.9 m, and
occupies a surface area of 2425 km2. Its catchment contains 3.7% of
the country’s population and 11.6% of its gross domestic product
(GDP) within its 36,900 km2 area (accounting for only 0.4% of
China’s total land area) (An and Wang, 2009). Historically, the lake
has served as a municipal water supply, irrigation source and
fishery and tourist destination, among other things (Qing, 2009).
Since the 1950s and especially since the 1980s, human activities
have placed a growing stress on the lake’s development. Rapid
industrialization and urbanization represent two of the most
important factors contributing to continually rising nutrient
concentrations in the lake. Lake eutrophication and human activi-
ties are destroying the healthy aquatic ecosystem of Taihu Lake,
disrupting the balance between algal and macrophyte abundance
(Ma et al., 2008). Our study area covers the entire Taihu Lake
surface based on the 1:50,000-scale Fundamental Geographic
Information of Jiangsu completed in the 1980s (Fig. 1). Land cover
types in the study area have experienced a significant change over
the past 30 years (Liu et al., 2007a). We categorized the land cover
types, which included upland (i.e. roads and buildings such as
docks, businesses and factories located along the shoreline), aquatic
vegetation and open water, according to the objectives of this study
and the field surveys conducted in 2009 and 2010. The aquatic
vegetation incorporated three types: (a) emergent vegetation,
including Phragmites communis, Zizania latifolia, Nelumbo nucifera
and others, with P. communis as the predominant species; (b)
floating-leaf and floating vegetation (referred to as floating-leaf
vegetation in later text although it includes both), including Eich-
hornia crassipes, Nymphoides peltatum, Lemna minor, Trapa bicornis
and others, with N. peltatum as the dominant species; and (c)
submerged vegetation, including Potamogeton malaianus, Cerato-
phyllum demersum, Hydrilla verticillata, Vallisneria spiralis and
others, with P. malaianus as the dominant species. Aquatic macro-
phytes, especially floating-leaf and submerged macrophytes, were
distributed primarily in the east and southeast of the lake, where
the water quality is significantly higher than in other parts of the
lake and where cyanobacterial blooms are rare (Lu et al., 2009).
Taihu Lake was divided into two parts for this study: the grass type
zone (i.e. the eastern and southeastern portions of Taihu Lake) and
the algae type zone (i.e. the remaining portions) (Lu et al., 2009).
2.2. Field surveys
Field surveys were carried out on 14e15 September 2009 and 27
September 2010. In 2009, a total of 412 ground-truth samples, of
100 D. Zhao et al. / Journal of Environmental Management 95 (2012) 98e107

Fig. 1. The study area showing the distribution of 903 training samples (412 in 2009 and 491 in 2010) of emergent, floating-leaf and submerged macrophytes and other types in
Taihu Lake. Sampling transects are represented by the blue (2009) and red (2010) lines. (For interpretation of the references to colour in this figure legend, the reader is referred to
the web version of this article.)

which 55, 98 and 116 samples represented emergent, floating-leaf
and submerged vegetation, respectively, were obtained from: (a)
208 plots located along a transect from the east to the south of the
lake (blue line in Fig. 1); (b) 137 plots from 26 lake locations
distributed nearly uniformly across the lake (Zhao et al., 2011); and
(c) 67 plots of reed vegetation or land (roads and buildings such as
docks, businesses and factories) selected from a 1:50,000 land use
and land cover map. Similarly, a total of 493 field samples, of which
60, 143 and 109 samples represented emergent, floating-leaf and
submerged vegetation, respectively, were obtained in 2010,
including 438 photographs taken along a transect from the east to
the southeast of the lake (red line in Fig. 1) and 55 plots from the
1:50,000 land use and land cover map. At each sampling plot,
photographs were taken. Plots were limited to areas where aquatic
vegetation was distributed uniformly across an area measuring at
least 60  60 m (i.e., four pixels of a Landsat ETM þ scene). The
camera was held at about 1.2 m above the water surface, with the
camera axis angled about 30 down from the horizon. The position
of each photograph was geo-located using a portable GPS receiver
with an accuracy of 3 m. In the laboratory, all the photographs were
interpreted visually and classified as emergent vegetation, floating-
leaf vegetation, submerged vegetation and open water (Fig. 2).
2.3. Image processing
Assessing and reducing the effects of clouds, cloud shadows and
haze is one of the most important preprocessing steps, especially for
aquatic remote sensing. Kloiber et al. (2002) recommended that
images with less than 10% cloud cover should be selected for the
estimation of water quality by remote sensing. However, this
criterion significantly reduces the number of suitable images and
can even lead to a complete lack of available images for important
growth seasons. This problem is especially common for the wet
summers (i.e. between June and September) in southern China.
Therefore, ETM þ images with less than 20% cloud cover were used
in this study. As a result, a total of 26 ETM þ images (path/row 119/
38 and 119/39, WRS-2 indexing system) covering 13 dates (i.e. 26
March, 29 May, 17 August, 2 September, 20 October, and 5 November
in 2009; 8 January, 13 March, 30 April, 19 July, 4 August, 20 August,
and 21 September in 2010) were selected for use in this study.
Using ERDAS IMAGINE 9.2 (Leica Geosystems Geospatial
Imaging, LIC), geometric correction was carried out for all the
images using second order polynomials with accuracy higher than
0.5 pixel, with a 1:50,000 land use and land cover map from 2010 as
reference. We removed pixels greatly contaminated by clouds and

Fig. 2. Examples of digital photos of the four cover types taken during field sampling. From left to right, cover classes are emergent vegetation, floating-leaf vegetation, submerged
vegetation and open water.
 D. Zhao et al. / Journal of Environmental Management 95 (2012) 98e107
cloud shadows. Because (a) no cloud filtering rules perform well in
all conditions (Aumann et al., 2003), (b) interactive interpretation is
one of the most stable methods (Di Girolamo and Wilson, 2003)
and (c) a limited number of images were used in this study, we
removed the cloud-contaminated pixels using interactive inter-
pretation. For the resulting images, atmospheric corrections were
made using the cosine approximation model (COST) described by
Chavez (1996), which has been used successfully in other aquatic
remote sensing studies for atmospheric corrections of multi-
temporal Landsat images (Chavez, 1996; Wu et al., 2007).
Given the similar spectral-radiometric responses of algal
blooms and aquatic macrophytes (Li et al., 2009), algal blooms
must be detected before aquatic macrophytes can be identified. In
order to detect algal blooms, we first detected the reed vegetation
in the algae type zone using the ETM þ image of 20 August 2010;
this image was taken soon after a rainfall event caused most of
the algae to sink, resulting in a near absence of algal blooms on
the water surface. We then identified pixels with a higher TM4
than TM3 reflectance, a simple method of distinguishing bloom
vs. non-bloom waters (Duan et al., 2008). For the grass type zone,
if a pixel of TM4 > TM3 appeared at least twice between August
and September, it was identified as aquatic macrophyte vegeta-
tion; if a pixel of TM4 > TM3 appeared only once during that
period, it was identified as a cyanobacterial bloom area. For the
algal bloom areas, the pixels of TM4 > TM3, except for reed
vegetation, were identified as cyanobacterial blooms. Our iden-
tification procedure was based on our field observations and
results reported in the literature: (a) the coverage of aquatic
macrophytes reaches its maximum with little variation between
August and September (Liu et al., 2007a), indicating a high
probability that pixels that are actually aquatic macrophytes will
occur at least twice in this period; (b) since almost no algal
blooms occur in the grass type zone and since cyanobacteria drift
with the wind (Lu et al., 2009), the probability that algal blooms
will appear twice in the same location is much lower; and (c)
almost no aquatic macrophyte vegetation occurs in the algae type
zone (Lu et al., 2009), except for some emergent vegetation
(mostly reeds) near the shore.
Mean reflectance values were extracted for ground-truth plots.
Multi-seasonal images have been determined to be more suitable
for aquatic vegetation classification than a single image (Ozesmi
and Bauer, 2002; Davranche et al., 2010). Therefore, in this study
we used a combination of two images, one winter and one summer.
We selected a total of three pairs of dates of ETM þ images, referred
to in later text as Image Pairs 1 (13 March and 21 September 2010),
2 (8 January and 20 August 2010) and 3 (26 March and 2 September
2009). Based on the reflectance of each plot, commonly used
multispectral indices and multi-seasonal indices were calculated
(Table 1). In this paper, the variables were calculated as the spectral
index values of a season or season combination. For example, NDVI-
(w), NDVI-(s), NDVI-(w-s) and NDVI-(s-w) are the NDVI of winter,
NDVI of summer, NDVI of winter minus NDVI of summer and NDVI
of summer minus NDVI of winter, respectively.
Table 1
Spectral indices used in this study.
Indices Abbreviation
Normalized Difference Vegetation Index NDVI
Simple Ratio SR
Normalized Difference Water Index NDWI
Normalized Difference Water Index of Mc Feeters NDWIF
Modified Normalized Difference Water Index MNDWI
Average reflectance of TM1, TM2 and TM3 AVE123
Simple Ratio for water clarity SRWC
101
2.4. Analytical methods
Classification tree (CT) analysis is based on dichotomous parti-
tioning of the data at certain thresholds of the values of the
explanatory variables, which determine the branch a particular
sample will follow. We used CT analysis (PASW-Statistics v. 18) to
develop a model for identification of aquatic vegetation. According
to aquatic vegetation characteristics, CT models were developed in
a progressive fashion (i.e., first emergent, then emergent þ floating-
leaf and then emergent þ floating-leaf þ submerged vegetation),
treating the remaining types as “other”.
We developed our initial CT models (Model-10) based on data
from the 2010 field survey and Image Pair 1. For comparison and
validation purposes, we also developed CT models based on data
from the 2009 field survey and Image Pair 3 (Model-09). We then
applied Model-10 to Image Pair 2 (different time period within the
same year) and Image Pair 3 (similar time period in a different year)
to determine the general reliability of CT models developed for one
time period when applied to a different time period.
To improve the general applicability of CT models to time
periods other than those for which they were created, we: (1)
developed traditional CT models by applying quantitative CT
analysis to images of a certain time period using ground-truth
samples; and (2) created new versions of the CT models with
modified thresholds that incorporated information concerning the
difference among the images of the different time periods. The
modified thresholds were defined as:
TSIm ¼ TSIo þ m (1)
where TSIo is the threshold from the original CT model developed
by applying quantitative analysis to images of a certain time period,
TSIm is the threshold modified for the new version of the CT models
developed when applying the original CT models to images of
another time period, and m is a modification coefficient. If a split-
ting rule selected for values greater than a threshold, the variable m
was calculated as:
m ¼ TSIm95%  TSIo95% (2)
If a splitting rule selected for values lower than a threshold, the
variable m was calculated as:
m ¼ TSIm5%  TSIo5% (3)
where TSIm-95%, TSIm-5%, TSIo-95% and TSIo-5% represent the average
values of the upper 95th and lower 5th percentile pixels in the
application images (subscript m) and the original images (subscript
o) of each spectral index (SI) used as a splitting variable (i.e., an
independent variable used to separate individual samples into one
of two branches ultimately leading to one of the dependent classes)
in the model. That is, after developing the initial classification trees
from the field samples and ETM þ image pairs as specified in the
previous paragraph and using these models to map the distribu-
tions of emergent, floating-leaf and submerged vegetation for the
Formula References
(TM4TM3)/(TM4 þ TM3) (Rouse et al., 1974)
TM3/TM4 (Pearson and Miller, 1972)
(TM4TM5)/(TM4 þ TM5) (Gao, 1996)
(TM2TM4)/(TM2 þ TM4) (McFeeters, 1996)
(TM2TM5)/(TM2 þ TM5) (Xu, 2006)
(TM1 þ TM2 þ TM3)/3 (Härmä et al., 2001)
TM3/TM1 (Kloiber et al., 2002)
102 D. Zhao et al. / Journal of Environmental Management 95 (2012) 98e107
different image pairs, we then extracted and averaged the SI values
of pixels with SI values in the upper 95th (or lower 5th) percentile
of all pixels classified as a particular aquatic vegetation type (based
on the CT splitting rules) from (a) the application image (i.e., the
images from a different time period to which the original CT was
applied) and (b) the classification image (i.e., the images used in
initial model development), resulting in the parameters designated
TSIm-95%, TSIm-5%, TSIo-95%, and TSIo-5%, respectively. For example, if
a rule for emergent vegetation designated a split where NDVI > 0.6,
then all pixels classified as emergent vegetation were listed in
descending order of NDVI and the pixels with NDVI values above
the 95% percentile were extracted and averaged. We called the final
classification trees that incorporated the modified thresholds CTm.
To determine possible effects of intrinsic and extrinsic factors
and explain the better performance of our modified-threshold
method over the traditional CT method, we modeled growth
dynamics of the three types of aquatic vegetation from winter to
summer using all the clear Modis images of the product MOD09GA
(from https://wist.echo.nasa.gov/api/). We used Modis images for
this temporal modeling because of its high temporal resolution (i.e.,
at least two daily images) relative to ETM þ data, for which the
temporal resolution is low (i.e., one image every 16 days). The
product provides MODIS band 1e7 daily surface reflectance at
500 m resolution and 1 km observation and geolocation statistics,
with necessary corrections for the effects of atmospheric gases and
aerosols. A total of 73 and 76 clear images of the entirety of Taihu
Lake were identified visually for 2009 and 2010, respectively, and
subsets of the images at typical areas in the lake were selected for
emergent, floating-leaf and submerged vegetation. The SI in Table 1
were calculated for each image and used to model the growth
dynamics of emergent, floating-leaf and submerged vegetation
from winter to summer (January 1eSeptember 30) with a logistic
model:
c 2010). This method was used for between-year comparisons. For
y ¼ þd (4)
1 þ ea,xþb
where y is the SI from Modis images of surface reflectance at 500 m
resolution, x is Julian day, d and c are the potential minimum and
range (maximum minus minimum), respectively, of SI values in
a year, and a and b are coefficients that were obtained using a least
squares algorithm defined by Zhang et al. (2003).
After the temporal dynamic of the SI was modeled using Eq. (3),
the difference between the calculated SI based on Modis data and
the modeled SI could be calculated. We used the magnitude of this
difference to quantify extrinsic factors affecting the SI. Because the
variation in the environmental or physical factors such as atmo-
spheric conditions, water quality and sun-view angle between
growth periods resulted in random deviation of calculated SI from
the simulation curve, these factors necessarily affected compara-
bility of the SI between growth periods and thus the applicability of
a CT developed for a certain time period to other time periods.
Contrasting with the variation in aquatic vegetation conditions
between time periods, the variation in these factors would alter the
SI value, and the magnitude of the change was used to quantify
extrinsic factors (EF) as follows:
1X
EF ¼ jSIc  SIm j (5)
n
where n is the sample number, and SIc and SIm are the calculated
and modeled SI values, respectively.
In a single growing season, the SI values also changed with the
development of aquatic vegetation. The variation in developmental
stage of the aquatic vegetation necessarily affected the compara-
bility of the SI between growth periods and thus the applicability of
CT models developed for a certain time period to other time
periods. Therefore, we quantified the variation of aquatic vegeta-
tion using the magnitude of the deviation in modeled SI from the
seasonal average. This magnitude was determined by aquatic
vegetation conditions and thus was considered to represent the
intrinsic factors (IF), which were calculated as follows:
1X
IF ¼ jSIm  SIm j (6)
n
where n is the sample number, SIm is the modeled SI and SIm is the
average SIm during a season.
EF and IF were calculated for two time periods, i.e. winter
(January 1eMarch 31) and summer (July 1eSeptember 31). More-
over, the significance of EF and IF was calculated using the training
samples from 2010:
jNTþEFþIF  NTEFIF j
SðEF þ IFÞ ¼  100% (7)
NT
where S(EF þ IF) is the relative significance of EF and IF in classi-
fying aquatic vegetation, and T is the relevant SI threshold from the
initial classification tree based on training samples. NT, NTþEFþIF and
NTEFIF are the numbers of samples classified as emergent,
floating-leaf or submerged vegetation when SI is greater (or less)
than T, T þ EF þ IF and TEFIF, respectively.
2.5. Model validation
Two methods were used to assess the stability of the CT models:
map accuracy based on ground-truth samples and comparisons of
distribution area of the vegetation types among the maps created
by the CT models. The former method uses a confusion matrix
calculated between ground-truth samples and classification results
and is a well-known and widely applied method (Davranche et al.,
image pairs within the same year, we opted for a cross-tabulation
between the classified pixels of each image pair, as this included
a much larger number of pixels and was thus more representative
of the changes in classification performance. We were able to use
this method because the area of aquatic vegetation in Taihu Lake
did not change much due to the vegetation being rooted and
therefore not drifting with the water currents, and because it
reached its maximum distribution between the end of August and
early September (Gu et al., 2005; Liu et al., 2007a; Lei et al., 2008).
Therefore, we compared area of distribution of the vegetation types
to assess the applicability of CT models developed for Image Pair 1
(13 March and 21 September 2010) to Image Pair 2 (8 January and
20 August 2010). Similar to the former method, we created
a confusion matrix of areas classified as the different vegetation
types by different CT models. Whereas the former method assessed
accuracy relative to actual ground-truth samples (hereafter termed
“accuracy”), the latter method assessed the degree of agreement
between classified maps created using the different CT models
(hereafter termed “agreement”).
3. Results
3.1. Models
Classification trees were developed for emergent, floating-leaf
and submerged vegetation by applying quantitative CT analysis to
the SI values obtained from Image Pair 1 (i.e. ETM þ for 13 March
and 21 September in 2010) using the vegetation type obtained from
2010 ground-truth samples. We called this set of models Model-10;
final model structures are shown in Fig. 3, with thresholds given in
 D. Zhao et al. / Journal of Environmental Management 95 (2012) 98e107 103

A B

Fig. 3. Classification tree model structures established for (A) emergent vegetation, (B) floating-leaf vegetation and (C) submerged vegetation. The end nodes are classified as:
3 ¼ emergent vegetation; 2 ¼ floating-leaf vegetation; 1 ¼ submerged vegetation; and 0 ¼ other. T1eT10 represent the quantitative thresholds for the classification trees as
determined by CT analysis or the method developed in this study (see Table 2).

Table 2 (see Model-10). Emergent, floating-leaf and submerged
vegetation were classified with accuracies of 91.4%, 96.5% and
88.1%, respectively (Table 3). The “other” category had a classifica-
tion accuracy of 98.9%. The overall accuracy for all four types was
94.9%. These results suggested that aquatic vegetation types in
Taihu Lake could be distinguished using classification trees if
adequate ground-truth samples were available. Therefore, we used
Model-10 to map the distribution of aquatic vegetation in Taihu
Lake in 2010 (Fig. 4); according to this map, emergent, floating-leaf
and submerged vegetation occupied 33.5 km2, 148.6 km2 and
129.6 km2, respectively, which amounted to 1.41%, 6.25% and 5.45%
of the entire lake area, respectively.
3.2. Model validation
To verify the stability of the developed CT models, we first
applied Model-10 to Image Pair 2 (i.e. ETM þ for 8 January and 20
August in 2010) and mapped aquatic vegetation distribution
accordingly. We then compared this map to that resulting from
applying Model-10 to Image Pair 1. Overall agreement (i.e., the
percentage of all pixels in the study area that were classified the
same in the two maps) was 92.5% (Table 4). However, the only class
with an agreement greater than 60% was the “other” class. There-
fore, in spite of the high overall agreement, this result was not
satisfying because our objective was to accurately identify specific
types of aquatic vegetation. For emergent, floating-leaf and
submerged vegetation, the classification accuracies were only
58.3%, 52.8% and 31.1%, respectively. Furthermore, only 24.0 km2,
90.3 km2 and 41.3 km2 of the lake area was classified as being
occupied by emergent, floating-leaf and submerged vegetation,
respectively, which constituted only 71.6%, 60.8% and 31.9% of the
area occupied by the respective vegetation type in the original map
(i.e., the map resulting from applying Model-10 to Image Pair 1;
Fig. 4). These results indicated that the stability of the CT for
different time periods within the same year needed improvement.
Second, using the same CT model structure as Model-10 (see
Fig. 3), we developed another set of CT models (Model-09) based on
2009 field samples and Image Pair 3 (i.e., ETM þ for 26 March and 2
September 2009) (see the thresholds for Model-09 in Table 2).
Because the image dates (Julian days) of Image Pairs 3 and 1 were
very similar, Image Pair 3 was deemed suitable for testing the
stability of CT in different years. Moreover, comparing the relative
abilities of Model-10 and Model-09 to classify aquatic vegetation
when applied to Image Pair 3 could be helpful in determining
potential reasons for misclassification. The thresholds of the split-
ting variables for Model-09 were substantially different from those
of Model-10. Additionally, similar to the results shown in Table 4,
the application of Model-10 to the 2009 images (i.e., Image Pair 3)
resulted in relatively low accuracy rates for emergent (69.0%),
floating-leaf (71.6%) and submerged (58.9%) vegetation (Fig. 5).
Overall accuracy was 78.7%. Compared with Model-10, Model-09
provided a substantially higher overall accuracy (i.e. 92.8%) when
applied to Image Pair 3. Moreover, the classification accuracies of
all the aquatic vegetation types were higher than 80%, i.e., 89.7%,
93.9% and 86.8% for emergent, floating-leaf and submerged
vegetation, respectively. These results further confirmed that CT

Table 2
Thresholds for the classification tree models (see Fig. 3) developed for classifying aquatic vegetation in Taihu Lake.

Models T1 T2 T3 T4 T5 T6 T7 T8 T9 T10
Model-10 0.422 0.236 0.569 0.375 0.962 0.129 0.052 0.126 0.296 0.054
Model-09 0.318 0.082 0.476 0.458 0.866 0.133 0.032 0.355 0.193 0.043
CTm-1 0.329 0.114 0.542 0.173 1.119 0.126 0.000 0.151 0.422 0.078
CTm-2 0.344 0.121 0.484 0.447 0.846 0.119 0.029 0.315 0.212 0.035

Model-10: CT models developed by applying quantitative CT analysis to Image Pair 1 using ground-truth samples from 2010.
Model-09: CT models developed by applying quantitative CT analysis to Image Pair 3 using ground-truth samples from 2009.
CTm-1: CTm version of Model-10 for Image Pair 2 in 2010.
CTm-1: CTm version of Model-10 for Image Pair 3 in 2009.
104 D. Zhao et al. / Journal of Environmental Management 95 (2012) 98e107

Table 3
Classification accuracy of Model-10 classification trees in identifying aquatic vegetation in Image Pair 1 (in number of field samples).

Emergent Floating-leaf Submerged Other types Classification

vegetation vegetation vegetation accuracy (%)
Emergent vegetation 53 3 1 1 91.4
Floating-leaf vegetation 0 138 2 3 96.5
Submerged vegetation 0 5 96 8 88.1
Other types 0 0 2 179 98.9

methodologies can be used to accurately identify aquatic vegeta-
tion types in Taihu Lake. The stability of CT could be improved by
changing and optimizing thresholds to incorporate differences
among time periods.
3.3. Performance of CTm
When applied to Image Pair 2, the CTm version of Model-10
based on Eqs. (1) and (2) and Image Pair 2 (i.e. Fig. 3 and CTm-
1 thresholds in Table 2) provided an overall map agreement of
96.2% when compared with the map resulting from application
of the original Model-10 to Image Pair 1 (see Fig. 4). For emer-
gent, floating-leaf and submerged vegetation, the classification
agreements were 86.5%, 81.2% and 72.6%, respectively (Table 5).
Total lake areas classified as each type were relatively similar to
those from Fig. 4 (i.e., 33.6 km2, 143.8 km2 and 118.9 km2,
respectively, or 100.3%, 96.8% and 91.7%, respectively, of the area
classified as these types in Fig. 4), and thus the result was quite
satisfactory.
Another set of CTm models was developed based on Eqs. (1) and
(2) and Image Pair 3 (i.e. Fig. 3 and CTm-2 thresholds in Table 2).
Compared with the original thresholds of Model-10, the modified
thresholds were much closer to those of Model-09 (Table 2) that
were obtained by applying quantitative CT analysis to Image Pair 3
and corresponding field samples. Overall map accuracy was
determined to be 90.5% (Table 6), which was also very close to that
of Model-09. For emergent, floating-leaf and submerged vegeta-
tion, the classification accuracies were 82.8%, 95.5%, and 82.1%,
respectively, differing only slightly from those of Model-09. These
results suggested that CT models with modified thresholds (i.e.
CTm), as developed in this study, were more stable than the tradi-
tional CT models.

Fig. 4. Map of the distribution of aquatic vegetation in Taihu Lake in 2010 developed by applying Model-10 to Image Pair 1.
 D. Zhao et al. / Journal of Environmental Management 95 (2012) 98e107 105

Table 4
Confusion matrix of area (km2) classified as the different vegetation types by Model-10 when it was applied to Image Pairs 1 and 2.

Image pair 1/Image pair 2 Emergent Floating-leaf Submerged Other types Classification
vegetation vegetation vegetation accuracy (%)
Emergent vegetation 19.52 5.25 0.02 8.69 58.3
Floating-leaf vegetation 3.08 78.46 0.80 66.24 52.8
Submerged vegetation 0.02 3.92 40.37 85.31 31.1
Other types 1.34 2.70 0.11 2055.59 99.8

Fig. 5. Classification accuracies of the maps obtained by applying both Model-10 and
Model-09 to Image Pair 3.
3.4. Intrinsic and extrinsic influences on model performance
The proportion of extrinsic factors (EF) relative to the sum of
extrinsic and intrinsic (IF) factors (see Eqs. (4) and (5)), i.e., EF/
(IF þ EF), ranged from 58.6% to 83.2%, with an average of 71.1%
(Table 7). This indicated that, on average, 71.1% of the instability of
the CT models when applied to different growth or time periods
than those for which they were developed was due to extrinsic
factors. The remaining (28.9%) instability was due to intrinsic
factors. For the ten spectral indices shown in Table 7, S(EF þ IF)
ranged from 5.2% to 394.4%, which suggested that the classification
accuracy of the CT models could be significantly influenced by the
actions of extrinsic and intrinsic factors.
4. Discussion
Our finding that the traditional CT models could not be applied
directly to images other than those for which they were developed
was not consistent with the published literature (Brown et al.,
2003; Baker et al., 2006; Davranche et al., 2010). This discrepancy
might be explained by environmental factors. First, at Taihu Lake,
atmospheric water vapor remained high for most of the year, and
especially in summer. Therefore, the criterion recommended by
Kloiber et al. (2002) regarding maximum cloud cover needed to be
relaxed in order to obtain sufficient remote sensing images. Some of
the summer images used in this study had higher cloud cover than
the Kloiber et al. (2002) recommendation. Although atmospheric
corrections were applied to the images, uncertainty in atmospheric
conditions still influenced the comparability of the remote sensing
signal from images of different time periods (Rokitnicki-Wojcik
et al., 2011) and thus also influenced the application of CT models
to different images. In the other studies mentioned above, it is
likely that the arid or semi-arid climatic conditions made this a less
serious problem (Brown et al., 2003; Baker et al., 2006; Davranche
et al., 2010). Second, the study areas in comparable studies have
typically been small lakes, reservoirs or vegetated wetlands (Brown
et al., 2003; Baker et al., 2006; Davranche et al., 2010); in contrast,
Taihu Lake is a large, shallow and windy lake. Water turbidity, one
of the most important factors reducing the applicability of CT
models between years (Davranche et al., 2010), varies with time
subjected to the action of wind (Liu et al., 2007a) and thus influ-
ences the applicability of CT models between images. Moreover, the
morphological adaptation of aquatic vegetation caused by water
turbidity can further affect the applicability of CT models between
images (Liu et al., 2007a, 2007b; Davranche et al., 2010).
Our results indicated that 71.1% and 28.9% of the instability of
traditional CT models when applied to different time periods
originated from extrinsic and intrinsic factors, respectively, con-
firming our suggestion, above, that the differential performances of
traditional CT models in different studies was the result of differ-
ences in environmental conditions. For the ten spectral indices
used in our CT models, S(EF þ IF) ranged from 5.2% to 394.4%, with
an average of 152.1%. This suggested that classification accuracy of
the CT models could be influenced substantially by the actions of
extrinsic and intrinsic factors. This is not surprising because the
reduced magnitude of the water signal in aquatic remote sensing
consequentially results in its being easily disturbed by environ-
mental factors such as atmospheric conditions, water quality and
sun-angle, which differs from terrestrial remote sensing (Silva et al.,
2008; Bustamante et al., 2009; Rokitnicki-Wojcik et al., 2011).
The calculation methodology and rationale of the CTm model
resulted in better performance of CTm over the traditional CT. The
threshold modification parameter was based on pixels for which
the SI values fell below the 5% percentile or above the 95%
percentile for a particular type of aquatic vegetation. These pixels

Table 5
Confusion matrix of area (km2) classified as the different vegetation types when Model-10 was applied to Image Pair 1 and the CTm version of Model-10 was applied to Image
Pair 2.

CTm/Model-10 Emergent Floating-leaf Submerged Other types Classification

vegetation vegetation vegetation accuracy (%)
Emergent vegetation 28.97 2.58 0.36 1.56 86.5
Floating-leaf vegetation 3.42 120.69 9.22 15.25 81.2
Submerged vegetation 0.02 14.41 94.12 21.08 72.6
Other types 1.17 6.11 15.21 2037.23 98.9
106 D. Zhao et al. / Journal of Environmental Management 95 (2012) 98e107

Table 6
Confusion matrix of field samples classified by Model-09 and the CTm version of Model-10 when they were both applied to Image Pair 3 (number of field samples).

CTm/Model-09 Emergent Floating-leaf Submerged Other types Classification

vegetation vegetation vegetation accuracy (%)
Emergent vegetation 24 1 0 4 82.8
Floating-leaf vegetation 0 63 0 3 95.5
Submerged vegetation 0 5 46 5 82.1
Other types 0 4 6 135 93.1

Note: Because of cloud contamination, only 296 of 412 samples were used.

Table 7
Quantification, from Modis images, of the extrinsic (EF) and intrinsic (IF) factors influencing the applicability of the original classification trees to different time periods. SI is the
spectral index, and S(EF þ IF) is the relative significance of EF and IF.

Aquatic vegetation SI R2a 2009 R2a 2010 EF IF EF/(IF þ EF) % EF þ IF S(EF þ IF) %
Emergent vegetation MNDWI-s 0.66 0.62 0.087 0.018 82.6 0.106 146.5
MNDWI-w 0.66 0.62 0.070 0.038 64.5 0.108 394.4
NDVI-s 0.77 0.79 0.095 0.019 83.2 0.114 115.8
Floating-leaf vegetation NDVI-(s-w)b 0.86 0.74 0.131 0.074 64.0 0.205 45.8
SR-s 0.41 0.53 0.393 0.132 74.7 0.526 120.4
B5-w 0.59 0.68 0.013 0.009 58.6 0.022 5.2
Submerged vegetation AVE123-(w-s)b 0.81 0.88 0.013 0.006 68.6 0.019 120.1
AVE123-s 0.81 0.88 0.019 0.007 73.2 0.026 181.3
NDVI-(s-w)b 0.79 0.92 0.113 0.048 70.1 0.161 193.2
MNDWI-s 0.72 0.89 0.179 0.074 70.8 0.253 198.5
a
Goodness of fit for the logistic model shown in Eq. (4) for Julian days between 1 January and 30 September; sample numbers were 50 in 2010 and 52 in 2009.
b
Average values of EF (IF) of NDVI-s and NDVI-w or AVE123-s and AVE123-w.

tend to be the most typical for each type of aquatic vegetation and
generally have the highest coverage or biomass; thus, they repre-
sent the vegetation type better than other pixels, and the difference
in the SI values between time periods for these pixels better
represents the influences of extrinsic and intrinsic factors. As
a result, modifying the original thresholds using the difference
between the extreme values eliminates much of the “noise” caused
by the intrinsic and extrinsic factors, improving the applicability of
CT between time periods. For example, the optimal thresholds (i.e.
the thresholds obtained using quantitative CT analysis with ground
samples) of AVE(123)-(w-s) for the classification of submerged
vegetation and “other” differed between Image Pairs 1 and 2, as did
the averages of the upper 95th percentiles due to the influences of
Fig. 6. Example demonstrating our explanation for the increased performance of CTm
with modified thresholds over the traditional CT with unmodified thresholds when
applied to images taken during different time periods. T1 and T2 are the optimal
thresholds (i.e. the thresholds obtained by quantitative CT analysis using ground-truth
samples) of AVE(123)-(w-s) for the classification of submerged vegetation and “other”
(not aquatic vegetation) obtained from 2010 field samples and ETM þ images (Image
Pairs 1 and 2, respectively). L1 and L2 represent the averages of the upper 95th
percentiles of AVE(123)-(w-s) from Image Pairs 1 and 2, respectively. D1 and D2 show
the differences between T1 and T2 and between L1 and L2, respectively. D1 approxi-
mates D2, which explains why the thresholds modified by extreme samples can
improve the stability of the CT models.
intrinsic and extrinsic factors (Fig. 6). However, the difference
between the extreme pixel averages approximated the difference
between the optimal thresholds; therefore, the original threshold
modified by the difference between the extreme value averages
was very close to the optimal threshold based on training samples.
This explains why CTm performs better than traditional CT with
unmodified thresholds.
5. Conclusions
Our results suggested that, for classification of aquatic vegeta-
tion types in Taihu Lake, traditional CT models developed for
certain image dates could not be applied directly to other time
periods within the same year or among different years without
modification. 71.1% of the instability of traditional CT models
applied to different time periods was due to extrinsic factors, i.e.
environmental and physical factors. We have shown that our CTm
models improved the stability of traditional CT models in mapping
the aquatic vegetation of Taihu Lake. This improved CT method-
ology will be useful for the investigation of aquatic vegetation and
lake management at large scales. For example, the ability to apply
the modified CTm models directly between growth periods enables
investigators to compile a complete mosaic of images from
different dates. This has special significance when clear images are
only partially available because of cloud contamination and
a mosaic is the only way to obtain complete coverage of the lake.
The ability to apply CTm models between years can be useful for
monitoring the spatio-temporal dynamics of aquatic vegetation,
and especially for obtaining information on past dynamics when
few or no field samples are available.
Acknowledgments
This study was supported by the State Key Development
Program for Basic Research of China (2008CB418201 and
2008CB418004).
 D. Zhao et al. / Journal of Environmental Management 95 (2012) 98e107 107

References Liu, W.L., Hu, W.P., Chen, Y.G., Gu, X.H., Hu, Z.X., Chen, Y.W., Ji, J., 2007a. Temporal
and spatial variation of aquatic macrophytes in west Taihu Lake. Acta Ecol.
Sinica 27, 159e170.
An, S., Wang, R.R., 2009. The human-induced driver on the development of Lake
Liu, W.L., Hu, W.P., Gu, X.H., 2007b. The biomass variation of Potamogeton malaianus
Taihu. In: Lee, Xuhui (Ed.), Lectures on China’s Environment. Yale School of
and its influential factors in Lake Taihu. Acta Ecol. Sinica 27, 3324e3333.
Forestry and Environmental Studies.
Lu, N., Hu, W.P., Deng, J.C., Zhai, S.H., Chen, X.M., Zhou, X.P., 2009. Spatial distri-
Aumann, H.H., Chahine, M.T., Gautier, C., Goldberg, M.D., Kalnay, E., Mcmillin, L.M.,
bution characteristics and ecological significance of alkaline phosphatase in
Revercomb, H., Rosenkranz, P.W., Smith, W.L., Staelin, D.H., 2003. AIRS/AMSU/
water column of Tahihu Lake. Environ. Sci. 30, 2898e2903.
HSB on the aqua mission: design, science objectives, data products, and pro-
Ma, R.H., Duan, H.T., Gu, X.H., Zhang, S.X., 2008. Detecting aquatic vegetation
cessing systems. Ieee. T. Geosci. Remote 41, 253e264.
changes in Taihu Lake, China using multi-temporal satellite imagery. Sensors 8,
Baker, C., Lawrence, R., Montagne, C., Patten, D., 2006. Mapping wetlands and
3988e4005.
riparian areas using Landsat ETM þ imagery and decision-tree-based models.
Macalister, C., Mahaxay, M., 2009. Mapping wetlands in the lower Mekong Basin for
Wetlands 26, 465e474.
wetland resource and conservation management using Landsat ETM images
Brown, E.C., Story, M.H., Thompson, C., Commisso, K., Smith, T.G., Irons, J.R., 2003.
and field survey data. J. Environ. Manage. 90, 2130e2137.
National park vegetation mapping using multitemporal Landsat 7 data and
Mcfeeters, S.K., 1996. The use of the normalized difference water index (NDWI) in
a decision tree classifier. Remote Sens. Environ. 85, 316e327.
the delineation of open water features. Int. J. Remote Sens. 17, 1425e1432.
Bustamante, J., Pacios, F., Díaz-Delgado, R., Aragonés, D., 2009. Predictive models of
Nelson, S.A.C., Cheruvelil, K.S., Soranno, P.A., 2006. Satellite remote sensing of
turbidity and water depth in the Doñana marshes using Landsat TM and
freshwater macrophytes and the influence of water clarity. Aquat. Bot. 85,
ETM þ images. J. Environ. Manage. 90, 2219e2225.
289e298.
Chavez, P.S., 1996. Image-based atmospheric corrections-revisited and improved.
Orth, R.J., Heck, K.L., Van Montfrans, J., 1984. Faunal communities in seagrass beds:
Photogrammetric Eng. Remote Sensing 62, 1025e1035.
a review of the influence of plant structure and prey characteristics on
Davranche, A., Lefebvre, G., Poulin, B., 2010. Wetland monitoring using classification
predator-prey relationships. Estuar. Coast 7, 339e350.
trees and SPOT-5 seasonal time series. Remote Sens. Environ. 114, 552e562.
Orth, R.J., Moore, K.A., 1983. Chesapeake Bay: an unprecedented decline in
Di Girolamo, L., Wilson, M.J., 2003. A first look at band-differenced angular signa-
submerged aquatic vegetation. Science 222, 51e53.
tures for cloud detection from MISR. Ieee. T. Geosci. Remote 41, 1730e1734.
Ozesmi, S.L., Bauer, M.E., 2002. Satellite remote sensing of wetlands. Wetlands Ecol.
Diaz, R.J., Solan, M., Valente, R.M., 2004. A review of approaches for classifying
Manage. 10, 381e402.
benthic habitats and evaluating habitat quality. J. Environ. Manage. 73,
Pearson, R.L., Miller, L.D., 1972. Remote Mapping of Standing Crop Biomass for
165e181.
Estimation of the Productivity of the Shortgrass Prairie, Pawnee National
Dogan, O.K., Akyurek, Z., Beklioglu, M., 2009. Identification and mapping of
Grasslands, Colorado. Processing of the 8th International Symposium on remote
submerged plants in a shallow lake using quickbird satellite data. J. Environ.
sensing of environment. ERIM, Ann Arbor, MI, pp. 1357e1381.
Manage. 90, 2138e2143.
Qing, B.Q., 2009. Progress and prospect on the ecodenvironmental research of Lake
Duan, H.T., Zhang, S., Zhang, Y., 2008. Cyanobacteria bloom monitoring with remote
Taihu. J. Lake Sci. 21, 445e455.
sensing in Lake Taihu. J. Lake Sci. 20, 145e152.
Rokitnicki-Wojcik, D., Wei, A., Chow-Fraser, P., 2011. Transferability of object-based
Fyfe, S.K., 2003. Spatial and temporal variation in spectral reflectance: are seagrass
rule sets for mapping coastal high marsh habitat among different regions in
species spectrally distinct? Limnol. Oceanogr 48, 464e479.
Georgian Bay, Canada. Wetlands Ecol. Manage.. doi:10.1007/s11273-11011-
Gao, B.C., 1996. NDWIea normalized difference water index for remote sensing of
19213-11277.
vegetation liquid water from space. Remote Sens. Environ. 58, 257e266.
Rouse, J.W., Haas, R.H., Schell, J.A., Deering, D.W., Harlan, J.C., 1974. Monitoring the
Graetz, R., 1987. Satellite remote sensing of Australian rangelands. Remote Sens.
Vernal Advancement of Retrogradation of Natural Vegetation. Type III, Final
Environ. 23, 313e331.
Report. NASA/GSFC, Greenbelt, MD, USA, pp. 309e317.
Gu, X., Zhang, S., Bai, X., Hu, W., Hu, Y., Wang, X., 2005. Evolution of community
Silva, T.S.F., Costa, M.P.F., Melack, J.M., Novo, E.M.L.M., 2008. Remote sensing of
structure of aquatic macrophytes in East Taihu Lake and its wetlands. Acta Ecol.
aquatic vegetation: theory and applications. Environ. Monit. Assess. 140,
Sinica 25, 1541e1548.
131e145.
Gullström, M., Lundén, B., Bodin, M., Kangwe, J., Ohman, M.C., Mtolera, M.S.P.,
Tan, B.X., Li, Z.Y., 2000. Rapid updating of rice map for local government using SAR
Björk, M., 2006. Assessment of changes in the seagrass-dominated submerged
data and GIS in Zengcheng County, Guangdong Province, China. Remote Sen.
vegetation of tropical Chwaka Bay (Zanzibar) using satellite remote sensing.
Land Resour. 43, 24e27.
Estuar. Coast. Shelf Sci. 67, 399e408.
Vis, C., Hudon, C., Carignan, R., 2003. An evaluation of approaches used to deter-
Härmä, P., Vepsäläinen, J., Hannonen, T., Pyhälahti, T., Kämäri, J., Kallio, K.,
mine the distribution and biomass of emergent and submerged aquatic
Eloheimo, K., Koponen, S., 2001. Detection of water quality using simulated
macrophytes over large spatial scales. Aquat. Bot. 77, 187e201.
satellite data and semi-empirical algorithms in Finland. Sci. Tot. Environ. 268,
Work, E.A., Gilmer, D.S., 1976. Utilization of satellite data for inventorying prairie
107e121.
ponds and potholes. Photogramm. Eng. Remote Sens 5, 685e694.
He, J., Gu, X.H., Liu, G.F., 2008. Aquatic macrophytes in East Lake Taihu and its
Wright, C., Gallant, A., 2007. Improved wetland remote sensing in Yellowstone
interaction with water environment. J. Lake Sci. 20, 790e795.
National Park using classification trees to combine TM imagery and ancillary
Jackson, J.B.C., Kirby, M.X., Berger, W.H., Bjorndal, K.A., Botsford, L.W., Bourque, B.J.,
environmental data. Remote Sens. Environ. 107, 582e605.
Bradbury, R.H., Cooke, R., Erlandson, J., Estes, J.A., 2001. Historical overfishing
Wu, G., De Leeuw, J., Skidmore, A.K., Prins, H.H.T., Liu, Y., 2007. Concurrent moni-
and the recent collapse of coastal ecosystems. Science 293, 629.
toring of vessels and water turbidity enhances the strength of evidence in
Jones, K., Lanthier, Y., Van Der Voet, P., Van Valkengoed, E., Taylor, D., Fernández-
remotely sensed dredging impact assessment. Water Res. 41, 3271e3280.
Prieto, D., 2009. Monitoring and assessment of wetlands using earth observa-
Xu, H., 2006. Modification of normalised difference water index (NDWI) to enhance
tion: the GlobWetland project. J. Environ. Manage. 90, 2154e2169.
open water features in remotely sensed imagery. Int. J. Remote Sens. 27,
Kloiber, S.M., Brezonik, P.L., Olmanson, L.G., Bauer, M.E., 2002. A procedure for
3025e3033.
regional lake water clarity assessment using Landsat multispectral data. Remote
Yang, Q., 1998. Ecological functions of aquatic vegetation in East Taihu Lake and its
Sens. Environ. 82, 38e47.
reasonable regulation. J. Lake Sci. 10, 67e72.
Lei, Z.X., Xu, D.L., Gu, J.G., Liu, Z.W., 2008. Distribution characteristics of aquatic
Zhang, X., Friedl, M.A., Schaaf, C.B., Strahler, A.H., Hodges, J.C.F., Gao, F., Reed, B.C.,
macrophytes and their Efects on the nutrients of water and Sedimnet in Taihu
Huete, A., 2003. Monitoring vegetation phenology using MODIS. Remote Sens.
Lake. J. Agro-Environ. Sci. 27, 698e704.
Environ. 84, 471e475.
Li, J.S., Wu, D., Wu, Y.F., Liu, H.X., Shen, Q., Zhang, H., 2009. Identification of algae-
Zhao, D.H., Cai, Y., Jiang, H., Xu, D.L., Zhang, W.G., An, S.Q., 2011. Estimation of water
bloom and aquatic macrophytes in Lake Taihu from in-situ measured spectra
clarity in Taihu Lake and surrounding rivers using Landsat imagery. Adv. Water
data. J. Lake Sci. 21, 215e222.
Resour. 34, 165e173.